How Sure Is That Accuracy?

Evaluate the same model on hundreds of random test sets and watch the measured accuracy wobble — why small benchmarks can't separate close models.